Multimodal Asset Generation
This challenge involves building an advanced generative AI system capable of producing creative marketing assets, including images with embedded text, based on complex briefs and brand guidelines. Leveraging the multimodal capabilities of Gemini 3 and Nano Banana Pro, participants will orchestrate a workflow that not only generates visually compelling images but also ensures accurate and contextually relevant text rendering directly within the image. The core of this challenge lies in integrating prompt optimization techniques using DSPy with sophisticated knowledge retrieval via LlamaIndex. This hybrid approach enables the system to dynamically adapt prompts for Gemini 3 and Nano Banana Pro, ensuring adherence to brand style guides and creative objectives fetched through RAG, while also self-correcting for improved text fidelity and image quality. This system will simulate a creative agency assistant, transforming abstract marketing concepts into concrete visual outputs.
What you are building
The core problem, expected build, and operating context for this challenge.
Build a generative AI system using Gemini 3 and Nano Banana Pro, DSPy, and LlamaIndex to create multimodal marketing assets with improved text rendering and adherence to brand guidelines.
How work is evaluated
The evaluation will assess the system's ability to generate high-quality multimodal assets that accurately follow creative briefs, adhere to brand guidelines, and exhibit excellent text rendering within the image. It will use a combination of automated metrics and qualitative assessment.
Challenge datasets
Public datasets, plus the private datasets you have access to.
What you should walk away with
Utilize LlamaIndex for Retrieval-Augmented Generation (RAG) to incorporate specific brand guidelines and creative examples.
Develop an evaluation mechanism for assessing text rendering quality and adherence to visual constraints within generated images.
Implement multimodal content generation using Gemini 2.5 Pro's capabilities.
Integrate DSPy for declarative prompt optimization and self-correction in a generative pipeline.
Reference links and supporting material
Sample data for 1 tasks
Challenge setup
The evaluation will assess the system's ability to generate high-quality multimodal assets that accurately follow creative briefs, adhere to brand guidelines, and exhibit excellent text rendering within the image. It...
Challenge input
{ "brief": "string", "brand_guidelines_doc": "string (path/URL to text doc)", "key_message": "string" }
Agent execution
No agent tool is listed for this challenge.
Evaluated output
{ "image_url": "string", "embedded_text_detected": "string", "metadata": {} }
- Verify that an image URL is returned and accessible.
- Confirm the key message is detectable and accurately rendered in the generated image (using OCR).
- Ensure LlamaIndex successfully retrieved and incorporated information from the brand guidelines (checked vi...
- TextReadabilityScore target: 0.9
- Python execution harness
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
No tool records are attached.
Evaluation contract
- Verify that an image URL is returned and accessible.
- Confirm the key message is detectable and accurately rendered in the generated image (using OCR).
- Ensure LlamaIndex successfully retrieved and incorporated information from the brand guidelines (checked via DSPy trace or agent logs).
Metrics
- TextReadabilityScore · Target 0.9
- CreativeBriefAdherence · Target 4.5
- VisualAestheticScore · Target 4
Workflow tasks
- Task 1
Input: { "brief": "string", "brand_guidelines_doc": "string (path/URL to text doc)", "key_message": "string" }
Output: { "image_url": "string", "embedded_text_detected": "string", "metadata": {} }
Execution harness: Python
Recipe freeze marker
No recipe freeze time is recorded.